AI Ed Wiki logoAI Ed Wiki
Use with AI

đź“„ Research Article

Control vs. Agency: Exploring the History of AI in Education

● high · created 2026-08-23 · updated 2026-08-23 · ai-education, history-of-aied, intelligent-tutoring, constructivist, agency, generative-ai

Synthesis: Mishra, Henriksen, Woo and Oster (2025) trace the history of AI in education from cybernetics and the 1956 Dartmouth conference through the cognitive revolution and Intelligent Tutoring Systems to today's GenAI, arguing that a seemingly mundane 1955 naming decision—John McCarthy choosing "artificial intelligence" over "cybernetics"—shaped the field's trajectory for decades. The paper frames the field's evolution as an "essential tension" between control and agency, embodied in the contrast between John Anderson's structured cognitive tutors and Seymour Papert's constructionist approach emphasizing creative learner agency. This tension recurs with each new technology wave and now resurfaces in debates about whether GenAI will reinforce traditional structures or promote greater learner agency and creativity.

Core Finding

The history of AI in education is not a story of technical inevitability but of contingent choices—most pivotally McCarthy's decision to name the new field "artificial intelligence" rather than "cybernetics." AI and education have been "inextricably intertwined since the 1940s," with early pioneers (Simon, Newell, Minsky, Anderson) investigating fundamental questions about learning and instruction. From this history two contrasting visions emerged: Anderson's cognitive tutors (AI to optimize instruction through precise modeling of knowledge and skills) and Papert's constructionism (computers as instruments of intellectual empowerment). These represent an "essential tension" (after Kuhn) between control and agency that manifests across knowledge and learning, roles and relationships, and political-institutional structures—and that recurs with each new technological wave, including GenAI today.

The 1955 naming decision: cybernetics vs. artificial intelligence

In the summer of 1955, as John McCarthy planned the Dartmouth Summer Research Project that would formally launch AI, he faced what seemed like a simple administrative choice: what to call the nascent field. The natural option was cybernetics, Norbert Wiener's term for the study of control and communication in machines and living things. But McCarthy rejected it for practical, not philosophical, reasons: "One of the reasons for inventing the term 'artificial intelligence' was to escape association with 'cybernetics.' Its concentration on analog feedback seemed misguided, and I wished to avoid having to accept Norbert Wiener as a guru or having to argue with him" (Nilsson, 2010, p. 78). While cybernetics emphasized systems, feedback loops, and interconnections, "artificial intelligence" framed machine capabilities in direct comparison to human cognition—an anthropomorphic framing that shaped research priorities, public perceptions, ethical debates, policy, and development priorities. The paper argues this choice, driven by institutional politics rather than philosophy, is now widely treated as natural and inevitable rather than a contingent historical decision.

The long, intertwined history of AI and education

The relationship between AI and educational theory is rich and multifaceted—from information processing to cognitive theories, from constructivism to constructionism. Even before the 1956 Dartmouth Workshop, Simon and Newell pursued questions that reshaped both domains: How do minds work? What is learning? How can we represent knowledge? Their research program advanced both information-processing psychology and symbolic AI, establishing fundamental links between human and machine cognition. The concept of human cognition as information processing—now fundamental to educational psychology—emerged from this work, epitomized in the cognitive revolution (Gardner, 1987). Newell outlined "12 aspects of intelligence... embodied in a meaningful teaching agent," advocating for automating "the essential intellectual operations involved" (Doroudi, 2023, p. 892).

This theoretical work led to practical applications: Intelligent Tutoring Systems (ITS), rooted in "heuristic search and expert systems paradigms of AI research of the 1960s and 1970s" (Kelkar, 2022, p. 20). ITS tracked student progress and generated targeted interventions, and were readily accepted in educational institutions because they aligned naturally with existing structures of assessment and standardization—the language of 'calculating learning interventions' reinforced traditional educational hierarchies and measurement-driven approaches. As the AIED community formed and funding priorities shifted, focus moved from foundational questions about intelligence and learning toward application-focused technical questions (optimizing learning models, refining adaptive algorithms), a shift that was "not merely technical but institutional."

Two visions: Anderson vs. Papert (control vs. agency)

Both Anderson and Papert built on the cognitive revolution but took radically different paths—one reinforcing traditional structures, the other challenging them.

Anderson's cognitive tutors (control): Anderson developed the ACT (Adaptive Control of Thought) theory, later ACT-R, structured as a series of production rules—a cognitive architecture designed to model human thought processes with precision and the foundational framework for his ITS. His Cognitive Tutors for LISP programming and geometry decomposed knowledge into explicit production rules, systematically presented instruction, and gave precise individualized feedback based on detailed models of cognitive processes, emphasizing learning-by-doing. Though Anderson's primary goal seemed to be validating his cognitive theory rather than building a learning tool, the tutors evolved into practical systems—and, crucially, his team shifted to align design with educators' priorities and curriculum standards rather than strictly following their theoretical framework. This adaptation revealed how educational technologies tend to "conform to existing structures rather than disrupt them."

Papert's constructionism (agency): Papert's vision emerged from his collaboration with Jean Piaget (who gave him insights about how children actively construct understanding—constructivism) and his AI research with Marvin Minsky at MIT (whose Society of Mind explored how intelligence emerges from interacting simple processes). Papert synthesized these into constructionism, which goes beyond Piaget's constructivism by emphasizing external artifacts—physical or digital objects that learners manipulate—as essential to learning. This was practically expressed in Logo, a programming language co-developed in 1967 with Wally Feurzeig and Cynthia Solomon, whose iconic 'turtle' interface created a 'microworld' where children could explore geometric concepts through programming. Logo embodied debugging-as-learning: finding and fixing errors as a natural, valuable part of learning rather than failure. Papert laid out this vision in Mindstorms (1980), arguing Logo was a medium for discovering powerful ideas across geometry, physics, and art—laying groundwork for Scratch, LEGO Mindstorms, and the maker movement, while challenging traditional education as "instructionism" that rejects children's natural learning processes.

Control vs. Agency: an essential tension

The tension between control and agency is what Kuhn called an "essential tension"—a fundamental, irreconcilable polarity shaping any domain concerned with human development. The contrast between ITS and constructionism manifests in three crucial dimensions:

  • Knowledge and learning: ITS breaks knowledge into discrete, manageable components with learning as step-by-step progression through sequenced material; constructionism embraces organic, project-based engagement where powerful ideas emerge from hands-on experience. ITS relies on formal production rules and explicit knowledge representation (enabling systematic assessment but feeling mechanical); constructionism emphasizes intuitive understanding developed through active engagement.
  • Roles and relationships: In ITS, the system serves as an intelligent guide providing precise feedback and interventions, with the teacher as a facilitator of this structured environment. Constructionism recasts both teacher and technology as enablers of discovery—teachers work alongside students as co-learners, and technology serves as a tool for investigation and creation rather than an instructor.
  • Political and institutional implications: ITS aligns naturally with traditional institutions' emphasis on standardization and assessment—a conservative vision prioritizing predetermined objectives, data collection, and algorithmic optimization, integrating seamlessly with learning analytics and personalization algorithms. Constructionism represents a more open-ended, learner-driven vision that resists easy quantification and challenges core assumptions about how learning should be organized and measured. This tension mirrors broader societal debates about control versus autonomy, standardization versus creativity, and whether education should serve existing structures or help transform them.

Control vs. Agency in the age of GenAI: personalization's two forms

Today's debates about GenAI in education closely mirror the historical ITS-vs-constructionism tension. In many cases, GenAI is positioned as a next-generation intelligent tutor—extending Anderson's ITS philosophy into the modern era with natural language dialogue, detailed explanations, targeted feedback, and adaptation to individual learning patterns. Central to this vision is the word 'personalization,' which has two distinct interpretations:

  • Uniform outcomes, varied path: adapts pace, difficulty, and teaching methods but ultimately guides all learners toward mastering the same predetermined curriculum. Rooted in Skinner's (1958) teaching machines and embedded in many ITS, this view is best captured by Sal Khan's TED Talk: one-on-one GenAI tutoring "could take your average student and turn them into an exceptional student."
  • Diverse outcomes: supports students in developing unique talents and pursuing individual directions, rather than diverse paths to similar outcomes. Zhao (2024) argues AI tools, unconstrained by traditional academic boundaries, enable students to pursue learning based on individual interests rather than conventional school subjects.

The ITS-as-tutor approach, while promising efficiency and accessibility, raises concerns about surveillance, data collection, and standardization, fitting comfortably within existing educational structures and corporate interests. A more transformative vision of GenAI offers possibilities for student empowerment and creativity but requires rethinking assessment, accepting ambiguous outcomes, and tolerating greater uncertainty—threatening established hierarchies. Who decides how these tools are used, and whose interests are served, are governance questions: the ITS model aligns with corporate and institutional interests in measuring, tracking, and optimizing learning, while the constructionist model suggests a more democratic and potentially disruptive approach.

Lessons for GenAI: the cybernetic counterfactual

Looking across the history, several patterns emerge: technical decisions consistently reflect deeper ideological positions; institutional forces repeatedly favor approaches that reinforce existing structures over transformative alternatives; and educators and researchers have consistently pushed back against dominant technological paradigms. The paper concludes by imagining the counterfactual—how different the educational technology landscape might look had McCarthy maintained Wiener's cybernetic framing. Instead of 'AI tutors,' we might speak of 'cybernetic learning systems' highlighting continuous feedback between learner and system; instead of 'personalized AI learning,' 'cybernetic adaptation' focusing on system-learner adjustment; the 'AI teaching assistant' might be a 'cybernetic learning mediator' facilitating connections rather than replacing human roles. These are not merely semantic differences: a cybernetic framing could emphasize interconnection over isolation, feedback over prediction, adaptation over optimization, and self-regulation over external control, steering away from the fantasy of machines replicating human intelligence toward technologies that support human learning in more systemic, ecological ways.

Relevance to the wiki

This is a landmark historical piece that anchors the wiki's developing history-of-aied theme, providing essential intellectual context for the many contemporary Intelligent Tutoring and Generative AI article entries. It offers the field's canonical origin story (the cybernetics-to-AI naming decision, the cognitive revolution) and its enduring conceptual framework (control vs. agency), which illuminates a wide range of wiki topics—from Personalized Learning's two forms to the pedagogical-agent-vs-tool debate, and from Adaptive Learning's alignment with institutional structures to constructionist tensions in AI Education. It connects directly to Learning Theories (constructivism/constructionism, information-processing), Agency (the essential tension), constructionism, Creativity, Teacher Role, and the ethics of AI in education. Its cybernetic counterfactual is a distinctive theoretical contribution useful across the wiki.

Connected Concepts

Connected Articles

Citation

Mishra, P., Henriksen, D., Woo, L. J., & Oster, N. (2025). Control vs. Agency: Exploring the History of AI in Education. TechTrends.

View SourceView ChangesEdit